Task 3.2: Understand lifecycle and limitations

• AAn entry for each of them, because every matching suggestion is logged as soon as it is shown

  Incorrect

  Entries are added when you accept a matching inline suggestion, not when one is shown.

• BNo entry for either, because references are only created for Copilot Chat responses

  Incorrect

  References are recorded for accepted inline suggestions as well as shown for chat responses.

• CAn entry only for the dismissed one, because rejected matches are logged for review

  Incorrect

  Dismissed suggestions never reach your code, and nothing is logged for them.

• DAn entry only for the accepted one, with details such as the matching file URLs and licenses

  Correct

  When you accept an inline suggestion that matches public code, Copilot records details such as the matching file URLs and any detected licenses.

For inline suggestions, code referencing compares a potential suggestion and about 150 characters of surrounding code with an index of public repositories. When you accept an inline suggestion that matches public code, Copilot records details such as the matching file URLs and any detected licenses, with the time you accepted the suggestion. Only accepted, unchanged suggestions are checked this way, so a suggestion you dismiss leaves no entry. Matches occur in less than one percent of suggestions.

• AHallucination

  Correct

  The cards define hallucination as output that sounds plausible but is incorrect, unsupported by the context, or fabricated.

• BContent filtering

  Incorrect

  Content filtering is a safety system that blocks harmful content.

• CPublic code matching

  Incorrect

  Public code matching checks suggestions against public code.

• DRed teaming

  Incorrect

  Red teaming is a testing practice GitHub carries out on its features, not a name for a kind of output.

Hallucination is a phenomenon where a language model generates output that sounds plausible but is factually incorrect, unsupported by the provided context, or fabricated. The inline suggestions card lists hallucination as a risk of pull request text completion and asks users to review generated text before publishing.

• ACompaction swapped old history for a summary that can drop fine details; /session checkpoints shows it

  Correct

  Compaction summarizes history and may lose fine-grained details, and checkpoints let you read what each compaction preserved.

• BThe session hit its token limit and is now dropping the oldest messages; /usage lists which ones were dropped

  Incorrect

  Compaction exists precisely so that the CLI does not drop messages without a record.

• CThe model changed during the session, so earlier messages were discarded; /model shows the switch

  Incorrect

  Changing models does not discard the conversation history.

• DCopilot Memory expired the fact after one hour; the repository's memory settings show it

  Incorrect

  Copilot Memory's retention is measured in days and is not what manages the session's context.

In long Copilot CLI sessions, compaction replaces old history with a structured summary, so fine-grained details such as exact wording or minor early decisions may not survive, and a completed compaction can't be reversed. Checkpoints keep each compaction summary; /session checkpoints lists them and /session checkpoints NUMBER shows one.

• AInput processing

  Incorrect

  Input processing is the first step: the prompt is combined with context.

• BLanguage model analysis

  Incorrect

  Analysis is the second step.

• CResponse generation

  Incorrect

  Generation is the third step.

• DOutput formatting

  Correct

  The response is finally formatted with syntax highlighting, indentation and links to context that was used.

The Copilot Chat application card describes the process as input processing (the prompt is combined with context such as the repository or open files), language model analysis, response generation, and output formatting with syntax highlighting, indentation and, depending on the question, links to the context the model used.

• ACode referencing only checks code from repositories with a license file

  Incorrect

  Matches include code with no detected license (shown as unknown or NOASSERTION).

• BCode referencing only runs on Copilot Pro+

  Incorrect

  The feature is not limited to a plan.

• CThe index is refreshed periodically and can miss recent code

  Correct

  The index is refreshed periodically and may not include recently added code.

• DCode referencing only checks private repositories in your organization

  Incorrect

  Private repositories and code outside GitHub are not included in the search.

Code referencing compares suggestions with an index of public repositories on GitHub.com; private repositories and code hosted outside GitHub are not included. The index is refreshed periodically, so it may not include recently added code and may contain references to code that has since moved or been deleted.

• AIt adds small trainable parts and leaves the base model unchanged

  Correct

  LoRA adds smaller trainable parts to each layer of the pretrained model instead of changing everything, saving time and resources.

• BIt retrains every weight of the network on the target dataset

  Incorrect

  That describes traditional full fine-tuning.

• CIt replaces the model with a smaller one trained only on GitHub code

  Incorrect

  LoRA adapts the pretrained model; it doesn't replace it.

• DIt removes layers so that suggestions are generated faster

  Incorrect

  LoRA adds trainable parts; it doesn't remove layers.

Fine-tuning tailors a pretrained LLM to a task with a smaller target dataset. Traditional full fine-tuning trains all parts of the network and is slow and resource-heavy. LoRA (Low-Rank Adaptation) adds smaller trainable parts to each layer while the original model remains the same, saving time and resources.

• ASessions can run for 24 hours, so the task must have hit a budget; raise the budget

  Incorrect

  The documented session limit is 59 minutes.

• BSessions run until the repository's Actions minutes run out; buy more minutes

  Incorrect

  The limit is a fixed maximum execution time, not an Actions quota.

• CSessions stop at 59 minutes; split the task into smaller ones

  Correct

  Each session has a maximum execution time of 59 minutes that can't be extended; break complex tasks into smaller ones.

• DSessions can be extended in copilot-setup-steps.yml up to 120 minutes

  Incorrect

  No setting extends a session: the maximum execution time can't be extended or bypassed.

Each Copilot cloud agent session has a maximum execution time of 59 minutes, which cannot be extended or bypassed; a task that exceeds it times out and stops. Break complex tasks into smaller, focused tasks.